How to build a chat bot product

    Co-founder of Roketbank, Viktor Lysenko, launched in Singapore a startup Osome, an online assistant for businesses. At the Epic Growth Conference conference , which was held in Moscow, Victor told how to build a chat bot product and what to focus on first.

    Watch the video and read the notes under the cut.

    Why chat-centric business can happen

    We all observe how the technologies of artificial intelligence are developing. In my opinion, in five to seven years, such a profession as an accountant who prepares documents for the payment of salaries or hiring employees is likely to either not exist at all, or the profession will turn into a very consulting function.

    From this follows the main value of the company. Osome automates a very large market - a large amount of manual labor. After all, until now, all accounting services are implemented manually. We decided to attack this market.

    Why chat is good

    Familiar interface

    The vast majority of people use the chat every day, using it as the main way to communicate.

    I send several hundred messages a day, and letters, probably within two dozen. Mail leaves. Chat comes. And, importantly, the user of the neural network was trained to perceive the chat interface. When you show the interface, no need to explain what to do about it.


    A customer wants to incorporate a company in Singapore or the United Kingdom, or in some other country. If he has never done this before, he will have many questions. What will he do first? Google Try to find articles where it is written about. The process is quite time consuming, because there are many specific cases. Like every person, he has his own special situation. Hard to find all the information at once.

    Let's see how the consultant acts: he first asks the client questions and, based on the responses received, leads the client to his case. With a large amount of choice and information for the client it is very convenient.

    Chats allow you to make wizards: ask a few questions to the user and lead to exactly the answer you need.

    One interface for many products

    If three or four years ago in chat rooms you could only send a text message, now widgets have appeared. This is a small piece of the interface, which has buttons, input fields, date selection, and so on. This makes it easy to direct the user and offer him a small piece of the interface right under his finger, that is, what he needs right now.

    This direction will be further developed in chat-bot-oriented products.

    Easier to replace people with robots

    Using chat as a communication channel makes it much easier to replace people with robots. At the current level of development of artificial intelligence, there are tools of Google, IBM, Facebook, Amazon, which allow you to determine the user's intentions and give him a sensible answer, which you prepared just for this question.

    If you compare with e-mail, then the chat is much more convenient. When a person writes a letter, he usually asks several questions at once on various topics. It is difficult to parse such a message and give the right answer to each of these questions. But with the messenger is much easier, because people usually write in short phrases. Ask one question. It is much easier to recognize and respond correctly.

    Chat - strategically one of the most successful solutions for automating a large volume of business, which is carried out manually.

    Why chat is bad

    Stupid bots

    Chat bots lead the user according to the developed scenario. If a person has an additional question or his situation does not correspond to the one that you, as a developer, have come up with in advance, he will leave.

    How to make bot smart and useful

    Chat bot is better than a person solves simple problems. And the person solves complex is much better. Therefore, people are friendly to chat bots in simple situations and are bad in complex ones.

    One of the main principles that we discovered for ourselves is that we did not fully automate, but combined bots and humans. We have an operating team, which has accountants and secretaries, they help answer questions. Here it is important that for the client, when he switches from robot to human and back - the process looks seamless, quick and clear.

    It is also one of the most important approaches: we always show customers who speaks with them - a bot or a person, and never give out a robot as a person and vice versa. We believe that it is necessary to be as frank as possible with our clients, and communicate with them as friends.

    Task driven vs Topic driven

    There are two types of dialogs: task driven and topic driven. Task driven - when the user has an ultimate goal, for example, to register a company. Topic driven - when a client wants to register a company and he has a number of questions: “How many shares can be issued?”, “And in what currency?”, “Do I need to get licenses?” And so on.

    That is, a person is studying, he has no end point to which you can lead him.

    The greatest difficulty is in combining the two types of dialogue. We are most interested in topic driven tasks that are converted to the ultimate business benefit. But along the way, the user can ask questions that relate to the features of the product or service.

    And the big problem is how to answer all his questions and then bring him back to the path we want us to take him on? This problem is, we have not yet found a good solution.

    Response rate

    Chat imposes a serious limitation on the speed of response. When the robot responds, there are no problems, since the robots respond instantly. In fact, you have to even slow down. The answer in the same second causes a negative perception on a subconscious person. It is necessary to show the timing indicator, as if someone writes, at least half a second.

    If the person answers - there are difficulties. It is difficult to configure the operation and admin panel so that the employee responds quickly. If a person solves the problem right now, try to fit the answers within a minute (while the user is ready to sit on the phone with an open chat).

    If not, then, in principle, anyway. We try to ensure that 80% of all answers come within the first 15 minutes. But ideally, good experience users get when you answer as in a real chat. As if on the other hand there is a person who is focused only on you and solves your problem.


    • It is important to engage people.
    • Integration with other systems.
    • Reduce the speed of response.
    • Do not build a complex system immediately.
    • Minimize cognitive load.
    • Speed ​​is important.

    We came to the conclusion: it is necessary to involve people in solving customer problems and in answering questions.

    Most likely, you will need integration with other systems. If you want to make your work useful, then you must have access to additional information. For example, in our case, this is access to the tax inspectorate base in Singapore. We take there the name of companies, identification numbers, the composition of shareholders and so on. It is very convenient for us and for the user.

    It is also important to try to reduce the response rate as much as possible. A large reservoir of the product manager’s work will be to make a back-office part of the system that will allow you to work and respond to customers as efficiently as possible.

    Do not try to build a complex system immediately. Choose an approach: short iterations. And look where users fall off, where conversion is worse or where customers start asking clarifying questions. This will make the product better.

    Customer development. Using the chat bot approach, you have a constant channel of communication with customers. You do not need a special way to communicate with customers, call them or make a mailing list. To find out how to improve the product, you can write directly to the chat and get answers very quickly. This gives a lot of feedback for quick product development.

    In October, we measured the NPS Osome. This number shows what percentage of customers recommend you to their friends. NPS = 88. For us, this means that the approach we have chosen works.

    More product marketing reports available on the @epicgrowth Telegram channel . Epic Growth Conference is a product marketing conference organized by Mobio, Getloyal and Appsflyer with the support of myTarget.

    The original full transcript of the speech was published on VC.

    Also popular now: